ChatGPT Guide

ChatGPT is OpenAI's AI platform: a chat interface for the GPT model family plus a growing set of tools around the chat (Projects, Skills, Apps, Workspace Agents, ChatGPT Work, Codex). As of May 2026, around 800 million people use ChatGPT every week; in the workspace segment the figure is around 15 million paid seats. Current model generation: GPT-5.6 (Sol, Terra, Luna), publicly available in ChatGPT since July 9, 2026. In the regular chat, Plus and higher use GPT-5.6 Sol, while Free and Go have used GPT-5.6 Luna with unlimited text chats since August 2026. In ChatGPT Work and Codex, GPT-5.6 is open to all plans. Since September 2026 the GPT-6 family has joined: GPT-6 Astra (September 3) as the top model, plus GPT-6 Sol and GPT-6 Luna (September 22) and GPT-6.1 Sol (September 29), these three for now only in ChatGPT Work and Codex.

1.1 The six building blocks

Model lock-in: Unlike Langdock, ChatGPT is tied to the OpenAI model family (GPT-6, GPT-5.6 and GPT-5.5 families, Codex, o-series). Anthropic's Claude models are not available in ChatGPT Enterprise. You choose the model per chat at the top left; the models on offer vary by plan.

1.2 Security, data protection, EU hosting

• Training opt-out: Active by default in Free, Plus and Pro (data is used for training; you can switch this off under Settings » Data Controls). Contractually excluded in Business, Enterprise and Edu.
• EU data residency: Available for Enterprise. Inference and storage in the EEA and Switzerland, with a dedicated region assignment per workspace.
• Compliance: SOC 2 Type II, ISO 27001, ISO 27018, CSA STAR Level 2. SAML and SCIM in Business, Enterprise and Edu. Audit logs via the Compliance API.
• EU AI Act: ChatGPT Enterprise meets Art. 4 (documenting AI literacy); a DSA point of contact is in place.

1.3 Plans at a glance

Details follow in chapter 3. Quick mapping (as of May 2026):

• Free and Go (~8 USD/month): personal use, low limits. Go available in 8 EU countries since January 2026 (AT, CZ, DK, NO, PL, PT, ES, SE).
• Plus (20 USD/month) and Pro (100 USD/month, new since April 9, 2026): personal power users, full feature set, no team workspace. The former Pro tier at 200 USD continues as "Pro 200" (previously called "Pro Max" here). New since September 29, 2026: Pro 500 at 500 USD/month.
• Business (20 USD/seat, reduced from 25 USD as of April 2, 2026): workspace for teams from 2 seats.
• Enterprise and Edu: custom pricing, from 150 seats, with SSO, SCIM, audit, Compliance API. ChatGPT for Healthcare as a HIPAA-safe Enterprise variant since January 2026.
• Teachers: free for US K-12 teachers until June 2027.

1.4 Who uses ChatGPT

As of Q1 2026, around 60 percent of Fortune 500 companies have at least one ChatGPT workspace. Among mid-sized companies in the DACH region, ChatGPT leads the adoption ranking: 47 percent of SME employees who use AI have used ChatGPT in the last 12 months (Bitkom 2025), often as shadow IT ahead of an official workspace rollout.

Typical use cases:

• Knowledge work: research, texts, analysis, code, briefing preparation.
• Content production: marketing copy, email drafting, newsletters, pitch decks.
• Customer support: reply drafts, knowledge base queries, FAQ maintenance.
• Data analysis: CSV evaluation, pivot tables, charts via Code Interpreter.
• Research agents: Deep Research for market and competitor briefings.

Productive in ten minutes. This chapter only shows you where things are. Content and use cases start in Ch. 4.

2.1 Where to access ChatGPT

2.2 Login

• Personal: account via chatgpt.com, login with Google, Apple or Microsoft. The free tier is available immediately.
• Workspace (Business, Enterprise, Edu): invitation by email or single sign-on (SAML, Microsoft Entra, Google, Okta). On first login, your personal account is linked to the workspace.
• Switching workspaces: In the top left corner there is a workspace switcher if you belong to several workspaces (e.g. a personal Plus and a Business from your company).

2.3 First tour of the interface

What you see after logging in:

• Left sidebar: chat history, Search, Library (with GPTs and Skills), Projects, Sora.
• Top left next to the logo: model selector (e.g. GPT-5.5 Auto, Pro, Thinking).
• Input field at the bottom: plus button for file upload, tool buttons for web search, Deep Research, image generation, Canvas, Code Interpreter, Voice button.
• Top right: account menu with Settings, Personalization, Memory.

2.4 What to do in the first ten minutes

1. Set language and tone: Settings, Personalization, Custom Instructions. An entry like "Always answer in German" is often enough (see Ch. 6).
2. Enable or disable Memory: Settings, Personalization, Memory. On by default. If you do not want ChatGPT to store facts about you, switch it off here.
3. Check data protection: Settings, Data Controls. Here you can disable "Improve the model for everyone" (relevant for Free, Plus and Pro).
4. Start your first conversation: "Help me with [specific task]". Start right away with a real use case instead of test questions.

ChatGPT has seven plans plus two special forms. The key dimensions: which models are available, how high the limits are, whether there is a workspace, and whether there are contractual training guarantees.

3.1 Plan comparison

Special forms:

• Teachers: Free for US K-12 teachers until June 2027 (not in DACH at present).
• Group Chats: Several users in one conversation, with ChatGPT as a participant. Being discontinued: no new Group Chats since July 9, 2026; existing ones will soon become read-only. More in Ch. 4.5.
• ChatGPT for Academic Researchers (since July 29, 2026): Researchers and postdocs can apply for 12 months of free access to a dedicated workspace. Up to five members per workspace, Business-level data protection and usage limits at Pro level. Process: sign up with a university email address, verify affiliation via SheerID, plus one qualifying publication and a short description of the research project. Then a 0 USD checkout with a valid card, create the workspace, invite up to four verified people from the same institution. At institutions with an existing ChatGPT Edu, approval runs through the institution. The program covers ChatGPT, not API credits.

3.2 Which model when

As of May 2026: GPT-5.5 Instant has been the default model in ChatGPT since May 5, 2026. 52.5% fewer hallucinations in law/medicine/finance, clearer answers, fewer superfluous emojis. GPT-5.5 Pro is the new top reasoning model. Update May 28, 2026: GPT-5.5 Instant and GPT-5.5 Thinking now deliver clearer answers and write code and text blocks directly in the chat. Canvas therefore no longer opens automatically in these two models (see Ch. 8). Update June 24, 2026: GPT-5.5 Instant has been refined for everyday conversations. The model is better at recognizing the actual goal behind a question, holds context more stably across several messages, follows complex multi-part instructions more reliably and adapts better when you follow up or disagree. Noticeable above all in decisions, advice, planning, research and shopping. Update July 9, 2026: The GPT-5.6 model generation (Sol, Terra, Luna) is publicly available in ChatGPT. In the regular chat, Plus, Pro, Business and Enterprise get GPT-5.6 Sol (via the effort levels Medium and higher); Pro and Enterprise additionally get GPT-5.6 Sol Pro for the highest quality. Free and Go initially stay on GPT-5.5 Instant in the chat. In ChatGPT Work and Codex, GPT-5.6 is open to all plans; free users get Terra there. Update August 6, 2026: GPT-5.6 Sol has been refined in the chat; it delivers more reliable facts and more focused, shorter answers. According to OpenAI, answers with factual errors occur around 68 percent less often than with GPT-5.5 Instant. New is a slider for thinking effort: for Plus and Pro, the same GPT-5.6 Sol drives both the fast answers and the deeper reasoning levels; switching up is meant to feel like the same model with more time, not like a switch to a different system. At the same time, GPT-5.6 Luna becomes the default for Free and Go, with unlimited text chats and a new Think button for harder questions (abuse protection still applies). Limits remain for file uploads, images and other tools. The update only affects the chat; ChatGPT Work and Codex are not changed by it. Retired on February 13, 2026: GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini, as well as the original GPT-5 Instant and GPT-5 Thinking. Retired on June 12, 2026: GPT-5.2 Instant, GPT-5.2 Thinking and GPT-5.2 Pro are no longer available in ChatGPT. Existing conversations that used GPT-5.2 automatically continue on the corresponding GPT-5.5 model. Retired on June 27, 2026: GPT-4.5 is no longer available in ChatGPT (not for Custom GPTs either). Existing conversations that used GPT-4.5 continue on GPT-5.5. Retired on August 26, 2026: OpenAI o3 is no longer available in ChatGPT. The 90-day transition period announced on May 28, 2026 has thus expired. Only ChatGPT is affected: in the API, the snapshots o3-2025-04-16 and o3-pro-2025-06-10 keep running until December 11, 2026. Update September 3, 2026: With GPT-6 Astra, the next model generation has arrived. It is stronger in coding, research, computer use and multi-step work, and it creates documents, spreadsheets and presentations that stick to given templates. The rollout started with a limited number of organizations and then continued across the paid plans, the API, Microsoft Azure and AWS Bedrock. Free and Go do not get Astra. Update September 14, 2026: OpenAI has announced the end of GPT-5.5. On October 14, 2026 the GPT-5.5 family disappears from ChatGPT, ChatGPT Work and Codex, on every plan: Free, Go, Plus, Pro, Business, Enterprise and Edu. The API is not affected. The successor is GPT-5.6 Sol. If you use Codex with a ChatGPT sign-in, switch from gpt-5.5 to gpt-5.6-sol beforehand. Check everything that hard-wires a model in good time: workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks and scripts. Update September 22, 2026: GPT-6 Sol and GPT-6 Luna are rolling out in ChatGPT Work and Codex at lower token prices than their GPT-5.6 predecessors. Sol is meant for complex and agentic work, Luna for focused, high-volume tasks. Available to Plus, Pro, Business, Enterprise and Edu, while Free and Go get Luna in the desktop app. Neither is available in regular Chat, only in Work and Codex. In Enterprise an admin must enable the new models first. Update September 29, 2026: GPT-6.1 Sol improves on GPT-6 Sol in agentic coding, computer use and professional work and, according to OpenAI, comes close to Astra at a much lower cost. OpenAI recommends it for repeated, long-running work across code, apps and documents. The rollout in ChatGPT Work and Codex starts with Pro and follows for Plus, Business, Enterprise and Edu. Enterprise and Edu owners enable the model under Workspace settings, Permissions & roles. Update October 7, 2026: GPT-6 arrives in the regular chat. In the model menu, GPT-6 replaces the previous Latest entry. Existing Latest selections move to GPT-6 at the same effort, while anyone who explicitly chose GPT-5.6 Sol stays there. For Plus, Pro, Business and Enterprise, GPT-6 Sol powers the Instant, Medium, High and Extra High levels, while the Pro level continues to run on GPT-6 Astra (GPT-6 Pro). Free and Go follow from October 8. New alongside it is Intelligent UI (box below). Update October 8, 2026: GPT-6.1 Sol is now also available as Ultrafast in ChatGPT Work and Codex, a faster service tier for Pro 500 and eligible Enterprise and Edu workspaces. In Enterprise, Ultrafast is off by default and owners must enable it. Ultrafast supports inference residency in the US and in Europe (EEA plus Switzerland). October 14, 2026: The GPT-5.5 family has been removed from ChatGPT, ChatGPT Work and Codex. For Codex signed in with ChatGPT, OpenAI now recommends GPT-6.1 Sol (gpt-6.1-sol) as the replacement. The API is not affected.

3.3 Limits and fair usage

• Free and Go: Since the rollout starting August 6, 2026, unlimited text chats with GPT-5.6 Luna, plus Think button for harder questions. Before that, it was 10 messages with GPT-5.5 Instant per 5 hours with fallback to a mini model. Limits still apply to file uploads, images and other tools, and abuse protection applies unchanged. Since August 14, 2026, the Think button for Free and Go has also been rolled out on the web.
• Plus: 80 prompts per 3 hours with GPT-5.5, separate limits for 5.5 Thinking and Pro.
• GPT-6 Astra (since September 3, 2026): Astra draws more budget per message than GPT-5.6 Sol, across all tiers roughly half the messages per five-hour window. For the highest Astra level, OpenAI states fixed values: Pro 200 so far 200 messages per week (from October 30, 2026: 100), Pro 50 per week, Business Premium 50 per week, Business Standard 15 per month. On Pro, GPT-6 Pro and GPT-5.6 Sol Pro share the same quota. Additional usage can be purchased as credits.
• Pro (100 USD): 5x Plus limits, Codex 10x as a promotion until May 31, 2026.
• Pro 200: 20x Plus limits in ChatGPT Work and Codex. Halved from October 30, 2026 to 10x, and the highest Astra level (GPT-6 Pro) drops from 200 to 100 messages per week. Existing subscribers keep the old limits until October 29 and receive a one-time usage credit of 2,500 USD, valid until the end of the year.
• Pro 500 (since September 29, 2026): Highest allowance of the Pro plans, OpenAI does not state an exact multiplier. Exclusive: Ultrafast for GPT-6 Astra in ChatGPT Work and Codex.
• Business and Enterprise: Workspace pool. Per-user limits similar to Plus, but central billing. Codex token pricing since April 23, 2026.
• ChatGPT Agent: No fixed monthly quota anymore. Agentic usage draws from a shared Agentic Pool, on the consumer plans as the tiers "limited" (Free, Go), "expanded" (Plus) and "maximum" (Pro), each with a 30-day reset. Business bills via included usage per seat plus workspace credits, Enterprise and Edu via a credit pool at contract level. Details in Ch. 18.4.

Current limits are listed in the official articles:

• ChatGPT Business: Models and Limits
• ChatGPT Enterprise and Edu: Models and Limits

This is where most people spend 90 percent of their ChatGPT time. This chapter shows you how to master the standard chat before you move on to Projects, Skills or Agents.

4.1 Anatomy of a chat

• Model selection: On the web directly in the input field (composer), on iOS and Android at the top of the conversation. Since June 10, 2026 the models are named by effort level (Instant, Medium, High, and depending on the plan Extra High, Pro Standard, Pro Extended) instead of the previous "Thinking" names. Set per conversation; if you switch mid-way, the history is kept.
• Model for a single message (iOS): On paid plans, long-press the send button to pick a different model for just this one message without changing the default.
• Tool buttons in the input field: Plus button for file upload, next to it buttons for web search, Deep Research, image generation, Canvas, Code Interpreter, Voice.
• Interactive charts (since June 8, 2026): ChatGPT can turn suitable answers directly in the chat into interactive bar, line, pie and scatter charts. Either you ask for one, or ChatGPT inserts one when it makes trends and comparisons easier to understand. (Web) Since August 26, 2026 these embedded visualizations in the iOS app follow the system's appearance and accent color.
• Table of contents (since June 8, 2026): Conversations with more than five answers get a table of contents that lets you skim sections and jump straight to them. (Web)
• Answer actions: Below each answer you find Copy, Regenerate, up- and down-vote, Read Aloud, share link.
• Branching: Clicking one of your own messages (pencil icon) allows "Edit and Resend". ChatGPT creates a new branch, the old history is preserved.
• Local time (since August 21, 2026): ChatGPT has a better sense of your local time in the conversation and takes it into account for time-related questions. Rolled out in the Android app from version 1.2026.216.
• Long chats load faster (since August 21, 2026): On the web, long conversations are loaded section by section instead of fetching the entire history at once. Since August 26, 2026 this also applies to the iOS app, where older parts of the history are only loaded on demand.
• Reasoning effort directly in the input field (iOS, since August 26, 2026): The iOS app shows a compact indicator in the composer where you can see and change the reasoning effort without switching to the model menu. Also new is a full-screen editor for long prompts. (App version 1.2026.230)
• Comments on specific parts of an answer (iOS, since August 26, 2026): You can optionally attach a comment to a highlighted passage in an answer instead of first describing your feedback in the next prompt.
• Interactive content appears step by step (since August 21, 2026): Interactive elements build up gradually in the conversation instead of only appearing once finished. For your own apps and widgets this means: they must be readable in a half-filled state too.

4.2 Chat history, search, pinning

• Sidebar on the left: Chronological history, grouped by Today, Yesterday, Last 7 days.
• Search (revised since July 14, 2026): Magnifier icon at the top of the sidebar. Search no longer covers conversations only, but chats, Projects, images and documents in one place. Filters let you narrow down by content type, and a click on a result opens the chat, the Project or the file directly. Available in all plans, on web, iOS and Android. Extension of August 7, 2026: Search additionally finds folders and also matches conversation titles (web, iOS, Android). On the web you can also search within the Library to find saved files.
• Pinning: Right-click a chat, "Pin". Fixes it at the top of the sidebar. A maximum of three pinned chats. Since August 20, 2026 pinned chats stay in sync between the desktop app and iOS, provided you use the same account there. Android is not yet included in this rollout.
• Scheduled tasks included in search (iOS, since August 26, 2026): Search also finds Scheduled Tasks, both by their title and by the content of the associated conversations. The task list now follows the desktop app in ordering and pinning.
• Folders: Workspace plans can file chats into folders.
• Archive: Archive old chats instead of deleting them. They disappear from the sidebar but remain searchable.

4.3 File and image upload

Via the plus button in the input field or drag and drop:

• Accepted formats: PDF, DOCX, PPTX, XLSX, CSV, TXT, MD, images (PNG, JPG), code files.
• Limits: 10 files per message, 512 MB per file (Plus, Pro), 2 MB for pure image files (Free).
• What happens to the files: They are read in the current conversation; in Projects they are stored in the Project knowledge. More in Ch. 12.
• Large text paste becomes an attachment (since June 22, 2026): If you paste a very large amount of text at once, ChatGPT automatically converts it into an attachment instead of overfilling the input field. This keeps the composer and the context window tidy. Threshold: from 10,000 characters. Since August 4, 2026 this also applies to Enterprise and Education, so the feature is now active in all plans (previously 5,000 characters). Via "Show in text field" you can bring the content back into the input field at any time.
• Formatting is preserved when pasting (since August 7, 2026): On the web, text you copy from Google Docs or from another ChatGPT conversation keeps its headings, bold formatting, links and lists, both while writing and after sending. No more manual reformatting.
• Saved files without re-uploading (since August 7, 2026): In the composer menu on the web, use "Add from library" to select a file you have already saved instead of uploading it a second time. See Ch. 12.2a.
• Answers closer to the file (since August 7, 2026): ChatGPT evaluates attached files more consistently when you have it summarize a document, look for a specific passage, check material or draft from a template.
• Audio uploads (since October 6, 2026): Upload audio files to create transcripts, summarize recordings or ask questions about their content, for example to turn a meeting, interview or lecture into structured notes or a follow-up draft. Only on paid plans and workspaces, depending on workspace settings, region, app version and model. Transcripts may contain errors, and quality varies by language.
• Scan documents (iOS, since October 1, 2026): In the ChatGPT camera, choose Scan from the three-dot menu and capture several pages in a row. ChatGPT automatically combines them into one PDF that goes straight into the chat.

4.4 Web search

With the search feature (globe icon) ChatGPT pulls live web sources. The answer format includes citation links. By default it runs automatically whenever ChatGPT decides that current information is needed (e.g. "Who won the game yesterday").

New since May 12, 2026: Free accounts now get more embedded images in search answers about well-known people, places and products. The images appear directly in the matching answer section and are clickable through to the original source.

4.5 Group Chats (being discontinued)

The feature launched on November 13, 2025 as a pilot in a few markets and was available globally from November 20. OpenAI is now withdrawing the pilot to keep the product leaner. How it works, as long as an existing Group Chat is still writable: several users in one conversation, with ChatGPT as a participant. All participants see the full history, write in real time, and ChatGPT answers context-aware.

What you should do now:

• Clarify in advance where things continue: If you still have an active group conversation running, decide where it will move to before the chat becomes read-only.
• Save a copy: If you need the history separately, request an export via ChatGPT or the OpenAI Privacy Portal.
• Replacement for the typical use cases: Collaborative brainstorming, joint research and live coaching in pairs will in future run in a team context via Projects (Ch. 7, shared projects in Business and Enterprise) instead of Group Chats.

4.6 Image generation in the chat (Images 2.5)

From May 2026, image generation ran on the ChatGPT Images 2.0 model. It can be called via the "Image Gen" tool button or by asking directly ("Generate an image of ..."). Available in all plans including Free.

• Fast mode: Immediate image output, good for mood boards and quick sketches.
• Images with Thinking: Plans and refines the image output before generating it. Higher quality, longer wait. Helpful for complex scenes, infographics or multi-element layouts.
• Edits within the image: Upload existing images, mark individual areas and have them changed specifically.

Limits: Free a few images per day, Plus considerably more, Pro practically unlimited. In Business and Enterprise, contractually no training on the prompts and images.

Update September 8, 2026: ChatGPT Images 2.5. Since September 8, 2026 image generation runs on Images 2.5, rolled out to all users in Chat, Work and Codex on web, desktop and mobile. According to OpenAI: sharper details, more natural lighting, more reliable multi-step editing and up to 50 percent shorter generation time compared to Images 2.0. Four new features in the user experience:

• Templates: Templates for common formats such as flyers or product photos as a starting point, instead of describing every layout from scratch.
• Sketch (mobile): Via @ and then Sketch you draw directly in ChatGPT. The sketch serves as a visual reference for the finished image. Useful whenever a spatial arrangement is easier to draw than to describe.
• Full-screen editing with comments: You place comments directly on a spot in the image instead of paraphrasing the desired change in the chat.
• Prompt sharing: You pass on the prompt you used so that others can reuse it with their own photos and details.

In the API this corresponds to two models: GPT-Image-2.5 Flare for speed and volume, GPT-Image-2.5 Sunburst for maximum editing precision.

Sticker packs (since August 31, 2026): From an idea or a photo you can generate a set of stickers, by asking for a sticker pack or via a sticker template on the home screen or in Images. On supported devices the stickers carry over to messaging apps, iMessage or WhatsApp.

4.7 Finances (US Pro, since May 15, 2026)

A personal finance experience on the web and in the iOS app. Pro accounts in the US can securely connect bank and credit card accounts via Plaid and get a dashboard that sorts expenses and answers questions relating to their own figures. As of May 2026, Finances is not available in DACH. No investment advice, no transactions, purely Q&A on your own data.

Update September 21, 2026: credit score. Plus and Pro users in the US can connect their Experian credit report including VantageScore 3.0. They get monthly updates, explanations of their score and alerts for new inquiries, new accounts or address changes. The Experian connection is separate from bank connections via Plaid. Not relevant for DACH.

Update October 2, 2026: Finances is now also rolling out to Free and Go users in the US, on the web, iOS and Android. Still not available in the DACH region.

4.8 Sharing a chat

Three options for passing on a conversation:

• Public link: Share button at the top right. Generates an anonymized link. Anyone with the link can read along, but not continue chatting.
• Workspace sharing: Within a Business or Enterprise workspace via "Share with team".
• Export: Settings, Data Controls, Export. Delivers a ZIP with all chats as JSON and HTML.

4.9 Job search and CV (since June 1, 2026)

ChatGPT can now search live for open positions and freelance jobs and help with your CV. It pulls current listings from sources such as Indeed, Upwork, Appcast and the open web, summarizes suitable matches and can create and format a CV based on your details. Available globally, in English for now. Useful as a research and writing aid; the application itself of course remains yours.

4.10 Health (US, since July 23, 2026)

A dedicated health area in ChatGPT, structured similarly to Finances. Logged-in Free, Go, Plus and Pro users in the US aged 18 and over can connect supported medical records and Apple Health data, see a dashboard with lab values, medications, activity and sleep, and can ask questions that draw on this personal context. Available on the web and on iOS, via the Health tab or directly in the chat, for example to understand trends, prepare for a doctor's appointment or make sense of test results.

• Data protection: Connected medical records, Apple Health information and the conversations that use them are, according to OpenAI, not used for training the foundation models and not for advertising. You control what is connected and when ChatGPT may access it.
• Limit: Health is explicitly a supplement, not a replacement for medical care, and is not intended for diagnosis or treatment.
• For DACH: As of August 2026 not available. Health data is a special category under Art. 9 GDPR, and OpenAI has not announced an EU release.

Addendum September 1, 2026: Healthcare Public Data. Eligible users of ChatGPT for Clinicians in the US can install a plugin that bundles nine public health sources, including PubMed, ClinicalTrials.gov, CMS Coverage, DailyMed and RxNorm, i.e. biomedical research, clinical trials, medication and Medicare data as well as provider registries. The apps are read-only and do not access patient records. Installation does not automatically connect all nine sources; each one is enabled individually. OpenAI explicitly points out that you should not write any health data of individual persons into searches sent to public sources. Currently without relevance for DACH, because it is tied to ChatGPT for Clinicians in the US.

Update September 14 and 28, 2026: New Health connections now follow the global plugin settings: low-risk actions run without asking, sensitive ones such as sharing data by email still need your approval (adjustable under Settings, Plugins, Health). Since September 28 you can select a chart, metric or record in the Health tab and get a personal explanation based on your connected data. Still not available in DACH.

4.11 Interactive quizzes (since August 14, 2026)

You can ask ChatGPT to quiz you on a topic and answer the questions directly in the conversation instead of receiving a wall of text with ten questions. A prompt such as "Quiz me on photosynthesis" generates a quiz card in the chat, with answer options, navigation between the questions and a hint button.

• Availability: all consumer plans (including Free) as well as Edu, on web and mobile.
• In practice: works well in combination with uploaded documents (Ch. 12) or a Project (Ch. 7), so that the questions come from your material rather than from general knowledge.

New since September 22, 2026: flashcards. On request, ChatGPT creates interactive flashcards on a topic or from uploaded notes. Tap a card to flip it, mark it as known or put it back for practice, and shuffle the order. The cards are saved automatically in your Library. Available on all plans on web and mobile.

4.12 Quick Chat with Pets and Appshots (desktop app, since September 11, 2026)

The desktop app for macOS and Windows has a floating control bar called Pets that sits above all other windows. You type a question into it directly and send it with Enter, without switching away from the program you are working in. Use @ to attach context and $ to pick a skill (Ch. 15). The bell icon lets you follow running threads and jump into the full conversation with a click.

• Showing it: Option+Space on macOS, Windows+Alt+P on Windows. Pressing again keeps Quick Chat open. You can change the binding under Settings, Keyboard shortcuts.
• Companion or just the bar: under Settings, Pets you pick a character or show only the controls.
• Appshots now on Windows too: pressing both Alt keys at the same time shares the frontmost application window with ChatGPT. On macOS you can start a chat in the Pets bar directly from an Appshot.
• Sources panel: files that appear as a source in an answer can now be opened straight from the panel.

4.13 Shopping: virtual try-on and favorites (since October 1, 2026)

Product listings for clothes and accessories now show a Try on button. Take or upload a selfie and ChatGPT Images generates a virtual try-on. The reference photo is saved for future try-ons and can be changed or deleted under Settings, Personalization, Reference photos. You can also save products to Favorites or collect them in folders in your ChatGPT Library. Available on web and mobile. Not relevant for company accounts, but useful to know when employees ask where their photo is stored.

A good prompt is not a trick and not a secret formula. It is a clear assignment with context, sources and expectations. With the GPT-5 models, the once important tricks ("You are an expert in ...") are less relevant, because the model itself makes good choices.

5.1 Four building blocks of a prompt

Microsoft calls it the GCSE pattern, OpenAI Academy speaks of the "Goal-Context-Source-Expectations" model. Both are the same thing:

• Goal: What should come out at the end? "Create a ...", "Summarize ...", "Compare ...".
• Context: Who is asking, in which role, with what background? "I am the sales lead at a DACH B2B company".
• Sources: What should ChatGPT access? File upload, "@Drive", web search, Memory.
• Expectations: Format, length, style, language. "Max. 200 words, 3 bullets, in German".

Goal: Write me the discovery call agenda for tomorrow.

Context: I am the sales lead at DECAID, a B2B AI enablement
company. My counterpart: managing director of a mid-sized company
(120 employees, mechanical engineering), first contact.

Sources: Look at the attached briefing PDF.

Expectations: 30-minute agenda, table with columns Time/Topic/
Question. In German. Polite, formal "Sie" form. No pricing statements.

5.2 When to use a more complex prompt structure

Rule of thumb: If you do the same task two or three times with the same setup, it does not belong in a prompt but in a Skill (Ch. 15) or a Project (Ch. 7). Prompts are for the individual case, Skills for repetition.

5.3 Anti-patterns

• "Be the best expert ...": No longer achieves anything today. The model decides for itself.
• Giant prompt with five tasks: Better two separate prompts. Otherwise ChatGPT loses track.
• Asking in English for German output: Keep it consistent. The language of the prompt influences the language of the answer.
• "Are you sure?": Does not help; ChatGPT will second-guess itself out of pure politeness. Instead: "Check the answer against [source]".

Custom Instructions and Memory are the two ways to tailor ChatGPT to you personally. Custom Instructions are an explicit brief, Memory is passive storage. Combined, they give you the biggest effect.

6.1 Custom Instructions

A global system prompt that applies to all conversations (except in Projects, which have their own instructions). Settings, Personalization, Custom Instructions.

Two fields. Since July 15, 2026 each holds up to 5,000 characters (previously 1,500) for Plus, Pro, Business, Enterprise and Edu. Free stays at 1,500 characters. More room does not automatically mean better: specific and concise beats long wish lists.

• What would you like ChatGPT to know about you? Role, industry, tools, context.
• How would you like ChatGPT to respond? Language, tone, format, length.

# About me
- Role: Marketing Lead at DECAID, B2B AI enablement DACH
- Tools: Notion, HubSpot, Figma, Slack
- Language: German, unless explicitly requested otherwise

# How should ChatGPT respond?
- No openers like "Sure!" or "Happy to!"
- Get straight to the point
- Structured with H2 or H3
- Max. 5 bullet points per list
- Cite sources explicitly
- Never "As an AI, I can't..."

6.2 Memory

Memory has two components:

• Saved Memories: Explicit notes that ChatGPT stores ("My name is Kiki", "My team has 12 people"). Editable under Settings, Personalization, Memory.
• Reference Chat History: Implicit learning from your history. Not a stored entry, but context enrichment in the background.

Both can be switched on and off separately. With Reference Chat History you cannot see directly what has been learned, which makes some users uneasy. Tip: If you want to know in a controlled way what ChatGPT stores, use only Saved Memories and switch Reference Chat History off.

Personalized suggestions on the home screen (since August 14, 2026): Eligible paid accounts see suggestions on the home screen that are derived from their own conversation history and connected tools. This is the same data basis as Reference Chat History: if you do not want the personalization, switch it off there.

6.3 Trusted Contact (Personal Safety, since May 7, 2026)

Optional safety feature for personal accounts: you designate a trusted person whom ChatGPT notifies in rare, serious suicide-risk situations so they can reach out. Set it up under Settings, Safety, Trusted Contact. Currently only in selected regions, not in Business/Enterprise/Edu accounts, and no substitute for emergency numbers.

6.4 Personality Presets

Since 2025 OpenAI has offered ready-made personalities (Cynic, Robot, Listener, Nerd). They complement Custom Instructions, they do not rewrite them. If you already have your own Custom Instructions persona, you will usually leave the presets on Default.

6.5 When to use what

6.6 Account security: Active Sessions and Lockdown Mode

Two new security settings for all logged-in accounts, both under Settings, Security:

• Active Sessions (since June 2, 2026): Overview of all active logins with device, location, login time and trusted status. Unknown sessions can be signed out remotely. Good for a quick check on whether someone else has access.
• Lockdown Mode (since June 4, 2026): Optional high-security mode that restricts access to the web and external services to reduce the risk of data exfiltration through prompt injection. While active, web search, Deep Research, Agent Mode, file downloads and some web-based images are blocked. Useful for accounts with particularly sensitive content.
• Security history (since September 25, 2026): Under Settings, Security and login, the security history shows recent sign-ins and sign-outs as well as changes to MFA, passkeys and other security settings, each with time, location and device. Some details are approximate or missing. A good first check when a sign-in looks suspicious.
• Privacy Center (since September 21, 2026): Brings together information on chat privacy, memory, personalization, data use, connected apps and account security, each with links to the relevant settings. On the web via the account menu, Help, Privacy Center, on mobile under Settings. For Free, Go, Plus, Pro and Business. The actual controls remain in Settings.

6.7 Computer History (macOS, since August 13, 2026)

Computer History is an optional feature of the macOS desktop app. When it is on, ChatGPT (and Codex) may draw on selected activity from your apps and websites as context, so you do not have to explain everything again in later conversations. Technically it evaluates interaction events via the macOS accessibility features, i.e. clicks, keyboard input, keyboard shortcuts and app switches, and periodically condenses them into text summaries and local memory files.

• Prerequisite: Memories must be active. The feature is not available with an API key or via Amazon Bedrock.
• Plans: Pro, Business and Enterprise in the macOS app. Off by default, each person has to opt in themselves.
• Admin approval: In Business and Enterprise, admins first have to enable the feature via Workspace Settings, Enable Computer History. Enabling only allows the opt-in, it does not switch anything on automatically (see Ch. 20).
• Switching on: Settings, Integrations. Via the menu bar icon, recording can be paused and resumed at any time, and individual apps and websites can be excluded.
• Reviewing and deleting: Settings, Computer History, History shows the activity grouped by day and time, with summaries and suggestions for automations. Deletion works individually or as a block for the last 10 minutes, the last hour, the last day, or everything.
• Retention: Temporary event files are kept for up to 48 hours. The memory files generated from them remain as files on your computer until you delete them.

Relevant for DACH: At launch on August 13, 2026, Computer History was not available in the EEA, Switzerland and the United Kingdom. Since August 20, 2026 the feature has been enabled there. The release notes explicitly name Pro for this, while OpenAI's own product documentation and the developer channel list Pro, Business and Enterprise for the supported regions. Business and Enterprise admins in the EEA should therefore check in their own workspace whether the toggle is already visible for them, instead of relying on either of the two statements. See also Ch. 21.

6.8 Temporary Chats with personalization (since August 27, 2026)

A Temporary Chat is the ephemeral conversation: it creates no new memories and does not appear in your history by default. What is new is that you decide at the start whether this chat may nevertheless access your existing Memories, Custom Instructions and Plugins, and that you can later move it into your history as a normal chat.

• Default stays unpersonalized: Without any action on your part, a Temporary Chat uses neither Memory nor Custom Instructions nor Plugins.
• Personalized: The chat reads your existing context but writes nothing back. No new memories are created.
• Saving: If something lasting does come out of the quick conversation after all, it can be saved as a regular chat.

Projects are ChatGPT's power-user tool. What Custom GPTs used to be, Projects are today: a persistent workspace with its own instructions, files, Memory and connectors. Project sharing via link has been available since Q4 2025, and since Q1 2026 Agent Mode also runs in Projects.

7.1 What a Project is

A Project bundles several chats, files and project-specific instructions into one container. All chats in the Project share the same context, the same Memory and the same files.

Examples:

• Weekly reporting Q2 2026: Files with OKRs, brand guideline, sample reports. Instructions: "Always write in Markdown, max. 1 page".
• Client X: Contract, communication history, style guide. Instructions: "Answer as a senior consultant, factual tone".
• Book manuscript: Outline, chapters so far, collection of quotes. Instructions: "Write in the first person, past tense".

7.2 Components of a Project

7.2a Changing the Memory setting afterwards (since August 14, 2026)

Until now, a Project's Memory setting was fixed when it was created. Now it can be changed at any time for non-shared Projects without creating a new Project: open the Project, three-dot menu, Project settings, under Memory choose between Default memory and Project-only memory.

• Project-only memory: ChatGPT uses only conversations from the same Project for context. Memories and chats from outside stay out, and conversely, information from the Project does not flow into the Memory that is used outside.
• Effect: The switch can take a few hours to take effect.
• Availability: all ChatGPT plans.
• Limits: Shared Projects stay permanently on Project-only memory and cannot be switched to Default. In Projects with Project-only memory, ChatGPT Work is not available.

7.3 Sharing

• Private (default): Only you can see the Project.
• Workspace sharing (Business, Enterprise, Edu): "Anyone at [your organization] with the link" or "Only those invited". Depending on their role, recipients can only read or also write.
• Plus, Pro, Free: Sharing is rolling out gradually during 2026. Currently workspace plans only.

Important: Recipients see instructions plus files plus the connector list. Check sensitive content before sharing.

7.4 When to use Projects, when Custom GPTs

7.5 Setup walkthrough

1. In the left sidebar, click "Projects", then "New Project".
2. Choose a name and color.
3. Optional: add instructions. Tip: specify persona, tone and output format.
4. Upload files via drag and drop or connect them via a connector.
5. Start the first conversation. All chats in the Project automatically share the context.

Canvas is a two-column editor. On the left the chat, on the right a document or code block that you both work on at the same time. ChatGPT edits directly in the document, you can highlight and comment or type yourself.

8.1 When to use Canvas, when normal chat

• Normal chat: Short answers, quick iteration, Q&A.
• Canvas: Longer texts, code, spec documents, essays. Anything that is revised across several responses.

8.2 How to use Canvas

1. Click the "Canvas" tool button in the input field or use the slash command /canvas.
2. ChatGPT opens the side-by-side view. The first version of the document is generated.
3. Highlight a section in the document, and a mini menu appears with "Edit", "Suggest Edits", "Add Emoji", "Adjust Length".
4. In the chat on the left you can give general instructions at the same time: "Make the whole text more formal", "Cut it down to 300 words".

8.3 Use cases

• Iterating a blog post: First draft in Canvas, then sharpen it paragraph by paragraph.
• Code refactoring: Open the function in Canvas, "Make this more idiomatic", then highlight individual spots specifically.
• Spec document: Outline in Canvas, then flesh it out piece by piece.
• Pitch email: First draft, then targeted "Sharpen the hook", "Make the CTA clearer".

8.4 Limits

• Canvas content becomes part of the chat. When you close it, it stays in the history, but not as a separate document.
• No real-time co-editing like Google Docs. If someone else is supposed to work on a Canvas in parallel, you have to share the content manually.
• Available in Free, Plus, Pro, Business, Enterprise, Edu (web, macOS, Windows).

Deep Research is a multi-step agent that researches the web on its own and returns a structured report with citations. 5 to 30 minutes of runtime, depending on depth. Available in Plus, Pro, Business, Enterprise and Edu.

9.1 When to use Deep Research

• Competitive briefing: "Who are Personio's five most important competitors in DACH and how do they position themselves?"
• Market research: "How big is the DACH market for AI enablement consulting in 2026?"
• Literature review: "What are the most important studies on AI adoption in mid-sized companies since 2024?"
• Due diligence: "What is publicly available about company X?"

9.2 Setting up a task

1. Click the "Deep Research" tool button in the input field.
2. Formulate the research brief. The more precise, the better. Tip: State not only the question, but also the goal of the research and the target audience.
3. ChatGPT shows a plan with the intended steps and source types. Confirm or adjust.
4. The research runs in the background. You get a push notification when it is done.
5. The result is a structured report in Canvas with inline citations.

9.3 What Deep Research can and cannot do

9.4 Limits

• Plus: 25 Deep Research tasks per month.
• Pro: 250 per month, longer runtime, several tasks at the same time.
• Business, Enterprise, Edu: Like Plus or Pro depending on the license, plus connector access.

9.5 Deep Research in Work and Codex (since September 10, 2026)

Deep Research used to be a feature of the normal chat. Since September 10, 2026 it is also available in ChatGPT Work and Codex, where it researches across the web, your own files and connected apps. The difference from the chat is less the research itself than the place: you start it in the same environment where the multi-step task is running anyway, adjust the scope during the run, and at the end you get an editable result with source references, not just a report to copy.

• What it is good for: When the research is meant to turn directly into a deliverable, for example a market screening of suppliers plus a finished overview, instead of reading a report and then building on it by hand.
• What stays the same: Runtime, source situation and limits from Ch. 9.3 do not change. Whatever is behind a login stays out, unless a connector brings it in.
• Prerequisite: Work is only available on paid plans, not in Free and Go (Ch. 18).

Code Interpreter (also "Advanced Data Analysis") is a Python sandbox in ChatGPT. You upload data, ChatGPT writes Python code, runs it, generates charts or new files. You do not have to write a single line of code yourself.

10.1 What Code Interpreter can do

• Data analysis: CSV, Excel, JSON. Pivots, aggregations, joins, correlations.
• Visualization: Charts (matplotlib, plotly), heatmaps, time series.
• File conversion: XLSX to CSV, PDF to text, scaling images, unpacking ZIPs.
• Statistics: T-tests, regression, clustering. Explains the result and the method in each case.
• Code snippets: Gives you the Python code to take with you.

10.2 How to use Code Interpreter

1. Upload a file via the plus button (CSV, XLSX, JSON, PDF, images).
2. Formulate the brief: "From this table, create a week-over-week comparison of revenue per store, visualize the top 5".
3. ChatGPT shows the Python code and the result. You can download both.
4. Refine iteratively: "Make the y-axis log-scaled", "Top 10 instead of top 5", "Add the comparison to the previous week".

10.3 Limits

• The sandbox resets per session. For longer-running analyses, it is best to create a Project, where files stay persistent.
• Files up to 512 MB per upload, max. 10 files in one conversation.
• Available in Plus, Pro, Business, Enterprise and Edu (not in Free).

ChatGPT talks to you, in real time, in many languages. Voice mode has been around as a feature since mid-2024. Since July 8, 2026, ChatGPT Voice runs on a new model generation: GPT-Live replaces the previous Advanced Voice Mode.

11.1 Standard Voice Mode

Classic speech output: you type, ChatGPT reads the answer aloud. Activated via the speaker icon below each answer. Available in all plans, all platforms.

11.2 ChatGPT Voice with GPT-Live

Real-time voice: you speak, ChatGPT answers immediately, you can interrupt at any time. Feels like a phone call. Activated via the voice icon at the bottom right of the input field. The rollout is global across iOS, Android and web.

• GPT-Live-1: Default for Go, Plus and Pro.
• GPT-Live-1 mini: Default for Free.
• Languages: Over 50, including German. Live translation during a conversation is possible.
• Voices: Several voice profiles to choose from (Settings, Voice).
• Limits: Plus around one hour per day, Pro significantly more. New since September 9, 2026: The daily limits for GPT-Live are unified across Go, Plus and Pro, and Plus and Pro no longer fall back to GPT-Live mini when the quota runs low.
• Model choice in Voice (since September 9, 2026): For search and harder questions, Voice can now draw on GPT-5.6 or GPT-6 Astra. You set the model and the thinking effort with the same controls as in text chat (Ch. 3.2). The former separate Voice intelligence levels Instant, Medium and High are discontinued as a result.
• Live Voice on the lock screen (since August 31, 2026, iPhone): Content from an ongoing Live Voice conversation appears on the Lock Screen and, on devices with Dynamic Island, there as well.
• Pronunciation help (since August 31, 2026): If you ask how to pronounce a word or phrase, you get tappable audio and a phonetic breakdown instead of just a transcription in the text.
• Files and Projects in a voice conversation (since August 7, 2026): GPT-Live in ChatGPT Voice now supports file uploads and Projects. You can upload a file in the middle of a voice conversation and have its content analyzed or queried, and use Voice inside a Project, including access to the most recent Project chats, the stored sources and the Project instructions.
• Protection for conversations in the background (since August 7, 2026, iOS and Android): If a voice conversation keeps running in the background, ChatGPT detects when it apparently is no longer being actively held, asks whether you are still there, and ends the session if there is no answer. Can be switched off under Settings, Voice, Background conversations.
• Smaller app improvements (since August 7, 2026): On iOS, ChatGPT automatically retries a dictation once after a connection drop or timeout instead of losing the recording, and finished answers appear immediately when you return to a chat instead of rebuilding themselves as if being generated. On Android, a voice conversation starts faster.

Update September 23, 2026: plugins in Voice, Voice in Work on web and mobile. Live now supports plugins and connected apps on web, iOS and Android, with the written response following along in the chat. Voice in Work is now also available on web and mobile: create documents, presentations and spreadsheets by voice, use connected apps and research the web. Anything unfinished when you hang up continues in text. Free and Go use Voice in Chat with the plugins their plan supports, Voice in Work requires access to both Voice and Work.

11.2.1 Dictation: new speech-to-text model

Independently of GPT-Live, at the end of June 2026 OpenAI swapped out the model behind the dictation feature, i.e. the conversion of your spoken input into text. The update applies to all plans including Free. Nothing changes in how you use it; it runs in the background.

• More accurate: The word error rate in the tested main languages is at least 10 percent below the previous model.
• Better with accents and mixed languages: Measurable gains for Japanese, Korean, Chinese, Urdu, Vietnamese, accented English, and for long dictations in Spanish, French, Italian and Portuguese. Also when multilingual speakers switch languages mid-sentence.
• More robust: In noisy environments, with quiet or whispered speech, and with combinations of letters and numbers, such as customer numbers or license plates.

11.3 Use cases

• Commuting by car: Have briefings read aloud, listen to research results.
• Language practice: Foreign languages with real-time correction.
• Hands-free coaching: Short consultations while cooking, exercising, walking.
• Brainstorming out loud: Instead of typing, just talk; ChatGPT helps you structure.

11.4 Realtime API (for developers)

If you want to build Voice into your own applications: OpenAI offers the Realtime API, GA since May 8, 2026, and since July 6, 2026 with model generation 2.1. Multimodal (audio in, audio out), low latency, suitable for voice bots, telephony and live translation. Not directly in ChatGPT, but via the API. Update September 10, 2026: GPT-Live 1 is generally available in the API and supports full-duplex voice conversations, i.e. speaking and listening at the same time, at a price of 0.05 USD per minute. This delivers on the previously announced API availability of GPT-Live.

Migration: The Realtime API beta was shut down on May 12, 2026. Anyone with code there has to switch to the final GA endpoints.

File upload is the fastest way to point ChatGPT at specific content. A PDF with the sales figures, a letter, an Excel file: upload, ask a question, and the answer comes back with reference to the content.

12.1 What you can upload

12.2 Limits

• Per message: 10 files.
• Per file: 512 MB (Plus, Pro), 2 MB for images in the Free plan.
• Per conversation: Effectively about 80 files in one chat history.
• Per Project: 25 files (Plus, Go), 40 (Pro), unlimited (Business).

12.2a File Library and storage quotas (since May 14, 2026)

The File Library brings together all files you have ever uploaded to or generated in ChatGPT in one place. Via the plus button in the composer you can access "Recent Files", browse everything in the "Library" sidebar entry, and ask ChatGPT directly about files you used earlier. Since May 14, 2026 also available for Free and Go accounts, including the EEA.

The quotas apply per account and are independent of the per-message upload limits. If you hit the limit, you can delete old files in the Library.

Update August 7, 2026 (web): The Library is more tightly integrated with chat. Via the composer menu and "Add from library" you attach an already stored file without uploading it again. In the Library itself you can select several files and folders at once by dragging across them with the mouse button held down. And search now also works within the Library, in addition to the global search across chats, Projects and folders (Ch. 4.2).

Update August 13, 2026: Google Drive in the Library. If you have connected the Google Drive plugin, you now see and search your Drive files and folders directly in the Library, including the items shared directly with you. You can also pull a Drive file into any chat via the composer or via @-mention, without uploading it again. Google Docs, Sheets and Slides can be kept open next to the conversation while ChatGPT summarizes, analyzes, compares or builds something new from them. You can also select an entire folder and let ChatGPT work across the files it contains. The content stays connected to Drive, and where supported and permitted, ChatGPT writes directly back to the source file.

Update September 10, 2026: Box, Dropbox and SharePoint are added. In addition to Google Drive, you can now also search and browse Box, Dropbox and SharePoint in the Library. You pull a file into the chat via Add from library or via @-mention, without uploading it again, keep it open next to the conversation, and let ChatGPT work across an entire folder as well. The rollout is running on the web for Go, Plus, Pro, Business, Edu, Healthcare and Enterprise, in Chat and in Work.

Update September 9, 2026: sharing files and folders. The Library is thus no longer just a personal storage place. You share individual files or entire folders with named people or with the entire workspace, assigning Viewer (read only) or Editor (add and maintain content), review and change the access later, and revoke it again. Shared items appear for recipients under Shared with me and can be used in conversations from there.

• Scope: initially "My Drive" plus the files and folders shared directly with you. Shared Drives are not included yet, and some editing and collaboration features of Google Drive are still missing in ChatGPT.
• Availability: Plus, Pro, Enterprise, Edu, Healthcare and Business on the web, in both Chat and Work mode. Mobile to follow.
• Caution when writing back: The fact that ChatGPT can change the source file directly is practical and risky at the same time. With shared documents, clarify beforehand who will see the change, and when in doubt have it work on a copy.

12.2b Space and Pages (since September 29, 2026)

For Pro, Business and Enterprise accounts on the web and in the desktop app, Space replaces Library. Space brings Pages, files and related work together in one place. You create a space for a project, add materials and share it with others.

• Pages: a new document format that teammates, ChatGPT and dots (ch. 18.6) edit together. You turn a conversation into a Page, start from a template or write directly in Space. Collaborators leave comments and each work with their own ChatGPT.
• Sharing: Sharing a Page does not grant access to private chats or memory.
• Projects stay: Projects (ch. 7) continue separately, for project chats, files and instructions.
• Announced: collaborative slides, exportable to PowerPoint and Google Slides.

12.3 What happens to the content

• In a chat: The file is processed in the current context and goes into the conversation's memory. No longer active after the chat is closed.
• In a Project: The file remains persistent. Every chat in the Project has access.
• In a Custom GPT: The file is stored in the GPT's knowledge. All users of the GPT access it.

12.4 Handling long documents smartly

With very long documents (50+ pages) the model becomes selective. Not every paragraph is read; instead, relevant parts are searched for. Three tips:

• Ask chapter by chapter: "In chapter 3, what is ..."
• Have it quote specifically: "Quote the passage where ...". Gives you the reference, which you can then check in the original.
• Summary with page numbers: "Summarize the main points with page numbers". That way you can verify things yourself.

12.5 Data protection when uploading

• Free, Plus, Pro: Content goes into the training pool unless you have disabled "Improve the model for everyone" in Settings, Data Controls.
• Business, Enterprise, Edu: Content is contractually excluded from training. Data processing agreement available.

Apps (known as "Connectors" until December 2025) connect ChatGPT to external tools. In the chat you can, for example, search Drive files, create GitHub issues or pull up HubSpot deals. As of 2026 there are around 30 native apps plus custom MCP servers for your own integrations.

13.1 What apps are

An app is an OAuth integration that gives ChatGPT read access, and sometimes write access, to an external service. When you say in the chat "Search Drive for Q2 forecast", ChatGPT calls the Drive API and returns the result with a citation.

There are two types of apps:

• Native apps: Integrations with the major platforms, maintained by OpenAI. No setup effort beyond the OAuth login.
• Custom MCP apps: Your own integrations that you or your team build using the Model Context Protocol. They also work in Claude and Codex (open standard).

13.2 Most important apps

13.2a Restaurant reservations (since August 10, 2026)

On request, ChatGPT searches for available restaurant tables directly in the chat, via OpenTable, Resy and Yelp. You describe where and when you want to eat, for how many people and what matters to you (cuisine, budget, dietary preferences, atmosphere). The available times appear directly in the reply, follow-up questions narrow things down further or check a specific restaurant, and clicking a suitable time takes you to the booking.

• Availability: Rolling out across all ChatGPT plans, on mobile, web and desktop.
• Regions: OpenTable worldwide, Resy in the US, Yelp in the US and Canada. For DACH this means only the OpenTable inventory is practically relevant.
• Not in ChatGPT Work: The reservation search is part of the chat, not of the Work agent.

Update August 28, 2026: multiple Google accounts at the same time. Until now, each Google plugin carried exactly one identity. Now multiple accounts for Gmail, Google Calendar and Google Contacts can be connected in parallel, so that a personal and a work account can be used in the same conversation. Set it up under Settings, Plugins, then for the respective plugin under Connected accounts click Connect another account. Available worldwide on Plus, Pro, Business and Enterprise, on web, desktop, iOS and Android. Conversely, for workspaces this means: a personal mailbox can be connected alongside the work one, which is something that belongs in your usage policy (Ch. 20).

Update August 21, 2026: plugin recommendations based on continued use. The ranking of suggested plugins is now based on which plugins people keep using after installing them, rather than on raw installation numbers. In practice this means the suggestions are more reliable, but they shift over time. If you maintain training materials with screenshots of the recommendation list, you should not treat it as a fixed list. Rolled out on web and mobile; desktop is not included. What you see at all still depends on plan, region and workspace settings.

13.3 Setup

1. Settings, Connected Apps. The list of available apps is displayed.
2. Select an app and click "Connect". The OAuth flow runs in the browser, and you confirm the permissions.
3. Call the app in the chat via "@"-mention or via the tool selector. Example: "@Drive search for Q2 forecast".

In the workspace plan, the admin can define which apps are enabled for which users.

13.4 Availability

• Free, Go: Selected apps available (Drive, GitHub, Gmail).
• Plus, Pro: Full app list.
• Business, Enterprise, Edu: Full list plus admin controls plus custom MCP apps.
• Plugins and marketplaces in the Admin console (since October 1, 2026): Owners and admins in Business and Enterprise now manage plugins and the workspace's plugin marketplaces in the Admin console (select the ChatGPT workspace, then Plugins or Marketplaces). Apps remain under Workspace settings, Apps, also reachable via Manage Legacy Apps. Since October 8, Enterprise and Edu workspaces with RBAC can set plugin access per role: workspace default Available, Install or Disabled, and custom roles inherit the default or get their own value.
• @ChatGPT in Slack and Microsoft Teams (since September 29, 2026): For Business and Enterprise. Via DM or a mention in enabled channels, ChatGPT answers questions, summarizes discussions or drafts text, using the company data and tools the admin has connected. Your organization must install and enable @ChatGPT first.
• Meetings plugin (beta, since September 29, 2026): In the macOS desktop app, it takes notes on what is said in a meeting and saves personalized summaries with action items in ChatGPT Space, private or shared with the team. The audio is deleted once the notes are ready. For Pro and Business, Enterprise to follow. Before using it in the DACH region, clarify consent from all participants and, where applicable, the works council.

13.5 Custom MCP apps

If you need your own integration (for example to your internal knowledge base), you can build an MCP server. MCP (Model Context Protocol) is the open standard developed by Anthropic and adopted by OpenAI at the end of 2025.

The advantage: an MCP server runs in ChatGPT, Claude, Codex and Cursor. So it is not OpenAI-specific. The effort for a custom integration pays off across tools.

Setup for workspace admins:

1. Build the MCP server (for example with the Anthropic SDK, Microsoft Agents Toolkit or OpenAI Apps SDK).
2. Add the server in the workspace settings.
3. Enable it for selected users or the whole org.

13.6 ChatGPT for Excel and Google Sheets (global since May 5, 2026)

A spreadsheet-native ChatGPT sidebar directly in Microsoft Excel (desktop and web) and Google Sheets. Installed as an add-in from the respective store. ChatGPT can build tables, explain formulas, clean data, rework multi-tab files and knows Skills and Apps (where available).

• Availability: Rolled out globally for ChatGPT Business. Plus and Pro in line with the Codex agentic limits, Free and Go with reduced limits.
• Free preview for Business: Until June 2, 2026 without counting against plan credits. After that, requests run through the normal usage quota.
• Data protection: In Business and Enterprise, workbook contents are contractually excluded from training.

ChatGPT for PowerPoint (GA since July 6, 2026): The same principle for presentations, available for Enterprise and Edu workspaces. ChatGPT creates and reworks editable decks directly in Microsoft PowerPoint, analyzes the structure and narrative of a presentation and draws on approved Skills and Apps while doing so. Polishing and template fidelity still need a review. Admins enable the add-in via the workspace settings or roll it out via Microsoft 365, including RBAC. For eligible workspaces, policy enforcement, scoped source access, the Compliance API, EU inference residency and Enterprise Key Management apply. Billing: For Enterprise customers, usage was free until August 6, 2026; since then it runs on the same token-based credit model as ChatGPT for Excel and Sheets.

ChatGPT for Word (since September 17, 2026): ChatGPT opens as a sidebar in Microsoft Word. You ask questions about the open document, turn notes into a first draft, summarize, revise selected text and adjust headings and formatting. It is installed through the same Microsoft add-in as for Excel and PowerPoint. Available on all plans including Free, usage counts against the normal plan limits. In Business workspaces, Word access is enabled by default since October 1, 2026, so anyone who does not want it has to switch it off actively.

Sign in with ChatGPT (since September 29, 2026): You use your ChatGPT account to sign in to third-party products without a separate password. Name, email and profile picture are shared. In a limited preview, Plus and Pro users can allow participating apps to use their ChatGPT model allowance, with app-specific limits. This usage counts against your own allowance, and the apps cannot see conversations or memory. In workspaces, organization settings apply.

13.7 Secure MCP Tunnel (new since May 19, 2026)

For Enterprise workspaces: private or on-prem MCP servers can be connected to ChatGPT, Codex, the Responses API and AgentKit without opening inbound firewall ports or exposing the servers to the public internet.

• How it works: A tunnel-client runs in the internal network, establishes an outbound HTTPS connection to the OpenAI tunnel control plane, fetches MCP requests, forwards them locally and returns the responses through the same tunnel.
• Enterprise networking: Outbound proxies, custom CA bundles, client certificates for the control plane and mTLS on the MCP side are supported.
• Use case: Internal knowledge bases, ERP or HR systems stay behind the firewall, and ChatGPT can still use them as an MCP tool.

Company Knowledge is an Enterprise feature: ChatGPT searches all connected company sources (SharePoint, Drive, Confluence, Slack, Notion etc.) and delivers an answer with citations. Like a company-wide search, only as a conversation.

14.1 What Company Knowledge does

Instead of addressing individual connectors ("@Drive ...", "@SharePoint ..."), an open question is enough. ChatGPT decides for itself which sources are relevant, summarizes and cites with a link to the source files.

Example: "What are our official brand colors?" Answer: "According to the 2026 brand guideline (SharePoint, Marketing folder), the colors are ..." with a link.

14.2 Availability

• Business, Enterprise, Edu: Available, admin-enabled. Currently ChatGPT web only (not yet in desktop or mobile).
• Plus, Pro, Free: Not available (workspace feature).

14.3 Current app connectors (May 2026)

Standard apps: Google Drive, Gmail, Google Calendar, Google Contacts, Outlook Email + Calendar, SharePoint, Dropbox, Box, GitHub, Linear, Notion, HubSpot, Teams. New in Enterprise/Edu since early 2026: Aha!, Asana, ClickUp, GitLab Issues, Help Scout, Teamwork.com, Zoho Desk, Azure Boards, Basecamp.

Plus: Japan data residency for all sync connectors since April 5, 2026. Microsoft Teams sync since May 15, 2026 for Enterprise/Edu.

14.3 Setup

1. The workspace admin connects the relevant connectors (Drive, SharePoint, Confluence, Slack, ...).
2. The admin grants access to individual user groups.
3. Users start a chat; "Company Knowledge" can be selected in the tool menu or is used automatically when the question relates to the company.
4. Answers come with a citation and a permission check, meaning: you only see sources you actually have access to.

14.4 What Company Knowledge is good for

• Onboarding: "Where do I find our travel expense policy?"
• Cross-team research: "What did Sales document about customer X last month?"
• Policy search: "What is our procedure for data breaches?"
• Unlocking the knowledge base: Instead of searching three SharePoint sites, ask once.

Skills are the most important new feature in ChatGPT 2026. Available in the ChatGPT workspace since April 2026, based on the open Anthropic standard. A Skill is a reusable workflow as a Markdown file with YAML frontmatter that ChatGPT activates automatically when a prompt matches its description.

15.1 What a Skill is

A Skill bundles three things:

• Trigger description: When should the Skill be activated?
• Instructions: What does the Skill do, and in what order?
• Optional assets: Templates, scripts, example files.

Unlike Custom Instructions (always active) and Custom GPTs (a separate bot), Skills are contextually active: they are only pulled in when their description matches the prompt. You can have dozens of Skills without them getting in each other's way.

15.2 SKILL.md format

A Skill is a folder with one mandatory file, SKILL.md, plus optional subfolders:

my-skill/
├── SKILL.md         (mandatory: frontmatter + body)
├── templates/       (optional: reusable templates)
├── scripts/         (optional: helper scripts)
├── references/      (optional: background documentation)
└── assets/          (optional: brand assets, example files)

The SKILL.md itself:

---
name: wochenreport-decaid
description: |
  Creates the DECAID weekly report in the standard format from
  OKRs, Slack highlights and calendar events.
  Trigger on "weekly report", "week status", "weekly review".
allowed-tools: ["web_search", "code_interpreter"]
---

# Weekly Report Skill

## Inputs
- Current calendar week
- Link to the OKR sheet
- Slack channel name

## Steps
1. Read the OKR sheet from references/okr-template.md
2. Aggregate Slack highlights (max 5 bullets)
3. Generate the report according to templates/wochenreport.md
4. Self-check via scripts/check_report.py

## Output format
H2 "Wins", 3 bullets
H2 "Risks", 3 bullets
H2 "Next week", 1 CTA

## Final Checks
- [ ] All four OKR areas mentioned?
- [ ] Max. 1 page?
- [ ] Risks clearly flagged?

Mandatory fields in the frontmatter:

• name: 1 to 64 characters, kebab-case.
• description: 1 to 1024 characters. Describes WHAT the Skill does and WHEN it is triggered. ChatGPT reads this text to decide whether to activate the Skill.

Optional fields:

• license: e.g. MIT.
• allowed-tools: List of the tools the Skill may use (web_search, code_interpreter etc.).

15.3 Setup in the ChatGPT workspace

1. Workspace sidebar, click "Skills" (only in Business, Enterprise, Edu).
2. "Add Skill". You can upload a folder or a ZIP file.
3. ChatGPT validates the frontmatter and shows a preview.
4. Activate the Skill, either for yourself only or for the whole workspace.

Workspace admins can control who is allowed to build, install or share Skills:

• Members create skills: yes or no.
• Members install skills: yes or no.
• Members publish skills for others: yes or no.

15.4 Cross-tool portability

This is the real lever: SKILL.md is an open standard. A Skill file runs 1:1 in:

• ChatGPT workspace (since April 2026)
• Claude Code and Claude apps (since December 2025)
• OpenAI Codex CLI (coding agent)
• Langdock (compatible format)
• Cursor, JetBrains Junie and many more (32+ tools in total)

15.5 Availability

• Business, Enterprise, Edu: Fully available, including admin controls.
• Plus, Pro: In beta rollout, enabled in most workspaces.
• Free, Go: Expected later.

Custom GPTs are pre-configured ChatGPT variants with their own name, logo, instructions, knowledge and conversation starters. OpenAI's hero feature in 2024, in maintenance mode since mid-2025: Projects have replaced them for power users. As of September 2026: the chapter is closing. Custom GPTs are being shut down and folded into Plugins, and from September 25, 2026 no new ones can be created. On personal accounts (Free, Go, Plus, Pro) creating and publishing is no longer possible anyway, while in Business, Enterprise and Edu workspaces it still works until the cut-off date, provided the workspace settings allow it. The rest of this chapter describes the estate you have to migrate before the shutdown.

16.1 When to use a Custom GPT

• Public sharing: You want to share a bot via the GPT Store or with a link to many people outside your organization.
• Conversation starters matter: Four pre-made entry prompts greet every new user. Projects do not have that.
• Stable workspace bots: "The HR bot", "The brand voice editor". Built once, used by everyone.
• Clear audience outside the builder: Other people are meant to use the bot, not edit it.

If none of these points applies, a Project (Ch. 7) or a Skill (Ch. 15) is the better choice.

16.2 Anatomy of a Custom GPT

16.3 Sharing options

• Private: Only you.
• Anyone with the link: Shareable via link, nobody but the recipient needed.
• Workspace: Within Business, Enterprise, Edu for all members.
• Public (GPT Store): In the OpenAI database. Searchable by category, star rating, number of users.

16.4 Setup walkthrough

1. In the left sidebar, click "GPTs", then "Create".
2. In the Builder tab, write a short description into the chat. ChatGPT generates a first draft.
3. Switch to the Configure tab and adjust the fields (Name, Description, Instructions, Conversation Starters).
4. Upload knowledge files.
5. Enable or disable capabilities.
6. Choose the sharing level, then "Publish".

OpenAI itself is thinning out its own GPTs as well: the official DALL·E GPT disappears on August 30, 2026, and image generation will run directly in the chat via ChatGPT Images from then on (details in Ch. 4.6). Self-built GPTs with image generation enabled are not affected.

Scheduled Tasks let ChatGPT run jobs at set times. You say "Send me a summary of last week every Monday at 8 a.m.", and ChatGPT takes care of it without your involvement and notifies you.

17.1 What tasks are

• Recurring: Daily, weekly, monthly or via a cron pattern.
• One-off: "Remind me on May 12 at 2 p.m. to send the briefing".
• Push and email: The result arrives as a push notification (mobile app) or as an email.

17.2 Setting up a task

1. Simply phrase it in the chat: "Schedule a task to send me a weekly summary of AI news every Monday at 8 AM".
2. ChatGPT confirms and shows the scheduled time.
3. Tasks can be managed in Settings, Tasks or at chatgpt.com/schedules. Status, paused tasks and the latest runs are listed in the Scheduled area.

Two forms. A standalone task starts a new chat for every run and reports the result under Scheduled, useful when each run should stand on its own. A task inside a chat returns to exactly that conversation and continues working with its existing context, useful for research loops or for keeping an eye on an ongoing process. Only tasks inside a chat allow minute-level intervals.

Your own rhythm. If daily, weekly and monthly are not enough, you set your own recurrence via the custom schedule control, in the advanced case directly as a recurrence rule per RFC 5545, for example RRULE:FREQ=MONTHLY;BYMONTHDAY=1;BYHOUR=9;BYMINUTE=0 for the first of every month at 9 a.m.

17.2a Triggering tasks by app events (since August 25, 2026)

Since August 25, 2026, a task no longer starts only at a set time but also when something happens in a connected app. Apps thereby turn from a lookup target into a trigger. The trigger determines when a run happens, the saved prompt determines what happens during it.

• Connect and authorize first: The app must be connected and authorized. For Slack, @ChatGPT must be a member of every monitored channel, otherwise the trigger does not fire. For GitHub, the connected app needs access to the respective repository.
• Combinable, but not with the clock: A task can have several event triggers, but event triggers and a schedule cannot be mixed.
• Web and mobile only: Event-triggered tasks are not available in the desktop app, the Codex CLI or the IDE extensions.
• Batching: If several matching events arrive in quick succession, ChatGPT may combine them into a single run. You can see pending events under Scheduled and process them there via Run now.
• Admin control: In managed workspaces, the permission Allow event-triggered scheduled tasks (under Permissions & roles) governs who may create such tasks at all.

17.3 Use cases

• Daily briefing: "Send me the most important tech news of the last 24 hours every morning at 7 a.m.".
• Weekly recap: "Every Friday at 5 p.m., take stock: what did I ask this week, which topics came up often?"
• Reminders: "Remind me on the 1st of every month to start the quarterly planning".
• Monitoring: "Check once a week whether a new pricing page is live on [competitor]'s website".
• Event-driven: "When an email from [customer] with "quote" in the subject arrives, extract the key data and create a summary for me". Here the tight filter matters more than the prompt.

17.4 Limits and availability

• Time-based tasks: Plus, Pro, Business, Enterprise and Edu, traditionally capped at 10 active tasks per account. Since August 2026, Free may create up to three active tasks with flexible daily recurrence.
• Event-triggered tasks: Plus, Pro, Business, Enterprise, Edu and ChatGPT for Healthcare. Free stays limited to purely time-based tasks.
• Rate limit for events: A maximum of 30 runs per hour and 720 per day, counted across all event-triggered tasks of an account combined.
• Team tasks (Business and Enterprise, since September 29, 2026): Recurring tasks can be shared within a team. They run on a schedule or supported trigger with approved company tools, and authorized teammates review results and update instructions together, so the task does not depend on one person. Example: a weekly project update in an approved Slack channel. The workspace admin sets up the connections, and a team owner enables them for the team.
• Sharing tasks (all plans): Scheduled tasks can be passed on. Recipients get their own copy and connect their own accounts, so a shared task does not carry the sender's app access with it.
• Model and effort: Both stay on default or are set explicitly per task. Tasks still running on gpt-5.4 or gpt-5.4-mini must be migrated by August 31, 2026: the replacements are gpt-5.6-terra and gpt-5.6-luna respectively.
• Desktop app: There, tasks also run project-based, either in the local project folder or in a separate worktree. This requires the computer to be on and the app to be open.

ChatGPT Agent is an autonomous agent for multi-step online tasks. Operator (operator.chatgpt.com) has been discontinued, all Operator functions are built into the Agent. The Agent alternates between reasoning and action, researches public websites, reads uploaded files and connector sources, fills out forms, edits spreadsheets and runs for minutes to hours. Enabled in Enterprise and Edu as well since April 2026.

18.1 What Agent can do

• Web research with action: Not just searching, but also filling out forms, clicking buttons, extracting data.
• Booking and appointments: Restaurant reservations, tickets, hotel bookings.
• Form filling: Application forms, contact forms, surveys.
• Data extraction without an API: Pull data from a web app that has no API.
• Multi-tab browsing: Several sites in parallel, comparisons, price research.
• Workspace Agents (since April 2026): Shared, always-on agents for teams with 60+ connectors, Slack integration, organization-wide permissions.

18.1a Agent tools (May 2026)

18.1b Browser integration: extension, Site Tools, cloud login (since August 25, 2026)

Since Atlas was shut down, the browser capability lives in ChatGPT itself. On August 25, 2026, OpenAI extended it in three places.

• Extension in five browsers: Besides Chrome, the ChatGPT browser extension is now also available for Microsoft Edge, Brave, Opera and Vivaldi. It lets ChatGPT read and operate pages you are already signed in to, such as LinkedIn, Salesforce, Gmail or internal tools. It is set up in the desktop app under Settings, Computer Use, after which you mention the browser in the chat with @.
• Site Tools (WebMCP): In the desktop app's built-in browser, ChatGPT Work and Codex can use tools that the website itself offers, instead of merely clicking through the page. This requires GPT-5.6 Sol or Terra and a current desktop app. With GPT-5.6 Luna and in Enterprise and Edu workspaces, Site Tools are not available.
• Login via the cloud browser: On eligible plans, ChatGPT Work on the web as well as on iOS and Android can ask you to sign in to a supported website. You enter the credentials in the login form, not in the chat. Cloud browser sessions stay separate from your local browser. This, too, is not available in Enterprise and Edu workspaces.

18.2 What Agent cannot (or may not) do

• Behind a login with your credentials: The Agent does not sign in on its own with your credentials, the model never gets to see them. If you need to get behind a login, use Connectors and Apps (Ch. 13), the browser extension in an already signed-in session or, since August 25, 2026, the website login via the cloud browser, where you type yourself (Ch. 18.1b). In Enterprise and Edu workspaces this login route is not available.
• Paying: Pause and user confirmation at every payment step. No autonomous purchase.
• Solving CAPTCHAs: At CAPTCHA walls, Agent stops and asks you.
• Sensitive sites: Banking and government portals are largely blocked.

18.3 Setup

1. Click the "Agent" tool button in the input field.
2. Phrase the job: "Compare the prices for [product] on the three largest DACH online shops".
3. Agent shows a plan with the intended steps. Confirm.
4. Agent runs in the background, you see a browser stream and can intervene at any time.
5. The result arrives as a summary with links and, where applicable, screenshots.

18.4 Availability and limits

• No fixed monthly quota anymore: The former hard numbers (Plus 40, Pro 400 agent messages per month) no longer apply. Agent, ChatGPT Work, Codex and Workspace Agents draw from a shared Agentic Pool. OpenAI deliberately no longer publishes a universal task count.
• Free and Go: Tier "limited", for looking rather than working. Plus: "expanded". Pro: "maximum". The allocation resets 30 days after first use in each case.
• Business, Enterprise, Edu: Business has included usage per seat plus a workspace credit pool, Enterprise and Edu have a shared credit pool at contract level without fixed rate limits. Enabled via workspace owner toggle (default OFF) and role-based enablement.
• Free, Go: Not available.
• Geo restriction: Currently available with restrictions in the EEA and Switzerland, individual functions blocked.

18.5 Security layers

• User confirmation for high-impact actions (purchases, sends, deletions).
• Watch Mode: Mandatory user supervision on sensitive sites.
• Refusal patterns for disallowed tasks.
• Prompt injection monitoring active.
• Blocklist of sensitive websites (banking, government) applies to the virtual browser and connectors.

18.6 Dots: always-on agents (since September 29, 2026)

A dot is a persistent agent in ChatGPT that stays on a task and keeps working between conversations. You give it a goal, connect the apps it needs and define what it may do on its own. Dots run on GPT-6 Astra with their own cloud computer, remember context and bring results back for your review.

• Availability: gradual rollout to eligible Pro and Business Premium users aged 18 and older in supported markets. You create your dot in the desktop app or on desktop web. Free and Go are excluded.
• Enterprise: as a beta, off by default, an admin must enable dots (ch. 20).
• Europe: Not available at launch for Pro accounts in the EEA, Switzerland and the UK (ch. 21).
• Cost: A first dot is included in Pro and Business Premium. During the first month, dot usage does not count against plan allowances, after that OpenAI plans to announce terms per plan.

Atlas is OpenAI's own browser, launched October 21, 2025. Chromium-based, with a built-in ChatGPT sidebar and integrated Agent Mode. As of May 2026: multi-account profiles, Saved Prompts, Tab Groups and a persistently extended Agent Mode. macOS GA, Windows remained in development until the end. The consolidation with ChatGPT Desktop and Codex announced for Q3 2026 has thus been decided: not as a merger, but as a shutdown of Atlas. The chapter remains as a description of what existed, so it stays clear what Atlas could do and where those capabilities live today.

19.1 What Atlas did differently (until August 9, 2026)

• ChatGPT always in the sidebar: While you browse, ChatGPT is displayed and knows the content of the current page. "Summarize this", "Quote the three most important points", without switching tabs.
• Agent Mode integrated: You can tell the browser "Order this on Amazon", and Atlas takes over using your existing logins. Since May 2026 noticeably more persistent on repetitive tasks (e.g. going through hundreds of emails).
• Browser Memories across tabs: Atlas extracts facts from visited pages; summaries are deleted after 7 days and can be switched off via user control.
• Multi-account profiles (since March 2026): Several ChatGPT accounts (Work/Personal/School), each with its own bookmarks, history, cookies and browser profiles.
• Saved Prompts (since April 2026): Save prompts via the bookmark icon, use them anywhere via @ or the bookmarks bar.
• Tab Groups + auto search mode: Vertical tabs, Tab Groups, "Auto" mode decides between a ChatGPT answer and a Google search.
• Proactive Suggestions (since April 2026): Personalized prompt suggestions next to "Ask ChatGPT" and above the composer.
• Smart Find (since May 2026): Cmd+F also finds semantically matching hits when there is no literal match.
• Privacy-first: More privacy defaults than Chrome, less tracking data trading.

19.2 Where Atlas makes sense

• Researchers and knowledge workers: Long sessions with many tabs, constant summarizing and comparing.
• Power shoppers and comparison testers: Agent Mode for price comparison and orders.
• Learners: Reading technical articles or documentation with ChatGPT as a tutor alongside.

Where Atlas does not fit:

• Banking, government services, sensitive logins: Better in the standard browser without an AI sidebar.
• Maximum performance: Atlas is somewhat heavier than bare Chrome.

19.3 Setup (historical)

Starting fresh is no longer possible; the browser only ran until August 9, 2026. The following steps document what the setup looked like.

1. Download Atlas from chatgpt.com/atlas (macOS, Windows; iOS and Android were announced but will no longer arrive).
2. Sign in with your ChatGPT account. From the first login, Memory and Custom Instructions are carried over.
3. Optional: import bookmarks and history from your old browser.
4. Show and hide the sidebar with Cmd+Shift+L (macOS) or Ctrl+Shift+L (Windows).

19.4 Availability

• Free, Plus, Pro: Fully available, all features.
• Business, Enterprise, Edu: The workspace admin can allow or block Atlas and set up compliance features.
• Geo: As of 2026 available in most countries, some features not yet active in the EEA.

19.5 Migration by August 9, 2026

This was the checklist for the migration. Anyone who did not work through it before the cutoff date can no longer access the browser data; ChatGPT conversations are unaffected by this:

1. Export bookmarks: Open the bookmark manager in Atlas and export as an HTML file. This file can be imported into Chrome, Edge or Firefox via their respective bookmark import.
2. Save open tabs and history: Does not carry over. Copy important URLs into a document beforehand or save them as bookmarks so they end up in the export.
3. Passwords and cookies: Are not migrated. If you saved logins in Atlas, you will need them again in the new browser, ideally from a password manager rather than from the browser store.
4. Choose a replacement: The ChatGPT sidebar effect can be recreated with the ChatGPT browser extension, which since August 25, 2026 is no longer available only for Chrome but also for Edge, Brave, Opera and Vivaldi (Side Chat everywhere except in Opera). The agent capabilities live in the ChatGPT desktop app and in ChatGPT Work (see Ch. 18.1b).

The new ChatGPT desktop app is gradually taking over what Atlas was meant for: multiple tabs, downloads, better navigation and login support for agentic tasks, depending on availability.

Workspace features are what make ChatGPT enterprise-ready: user management, SSO, audit logs, compliance evidence, data residency. Anyone rolling out ChatGPT in an organization should know the key levers, even if the day-to-day work is covered in another chapter.

20.1 Roles in the workspace

20.2 Single sign-on and SCIM

• SSO: SAML 2.0. Supports Microsoft Entra (Azure AD), Google Workspace, Okta, OneLogin, Auth0, JumpCloud.
• SCIM: Automatic user provisioning and deprovisioning. When a user is deactivated in the identity provider, they immediately lose ChatGPT access.
• Domain Verification: Before SSO is activated, the domain is verified via a DNS record. This prevents private accounts with the company email from existing outside the workspace.

Setup in the Global Admin Console (admin.openai.com), Identity section. Not active by default, must be set up explicitly.

20.3 Audit logs and Compliance API

• Audit logs: Who performed which action and when. Login, workspace change, GPT creation, file upload, project sharing.
• Compliance API: Programmatic access to conversations and logs. For SIEM integration and eDiscovery. Available in Enterprise.
• Workspace Analytics: Activity dashboard with usage figures per user, department, tool. Helps track adoption.

20.4 Data Controls

What OpenAI does with workspace data is governed by the Data Processing Addendum:

• No training: Contractually excluded in Business, Enterprise and Edu.
• Storage: Conversations are stored for service operation. Zero retention can be configured on request (Enterprise).
• User deletion: Right to be forgotten via the workspace admin or via API.
• Memory obligations: The admin can switch Memory on or off tenant-wide.
• Computer History (since August 13, 2026): The macOS feature that evaluates activity from apps and websites as context can only be activated at all in Business and Enterprise after admin approval (Workspace Settings, Enable Computer History). The approval does not switch anything on; it merely allows the individual person to opt in. Because this creates local, unencrypted memory files on the end devices, the decision belongs in the usage policy and not in the individual case. Details in Ch. 6.7.
• GPT-6 Astra (since September 3, 2026): In Enterprise, the new model is switched off at launch. An admin must enable it for the workspace, otherwise members won't see it at all. So if you search for Astra in the workspace and can't find it, it's usually not a rollout issue but an approval issue (Ch. 3.2).
• GPT-6 Sol, GPT-6 Luna and GPT-6.1 Sol (since September 22 and 29, 2026): These models, too, only appear in Enterprise and Edu after approval. Owners enable GPT-6.1 Sol under Workspace settings, Permissions & roles, on the Workspace tab or through a custom role.
• GPT-6 in Chat and Ultrafast (since October 7 and 8, 2026): GPT-6 in Enterprise Chat uses the existing GPT-6 Sol approval, there is no new admin preview, and admins can turn access off. GPT-6.1 Sol Ultrafast is off by default in Enterprise and must be enabled by owners.
• Dots (since September 29, 2026): A beta in Enterprise and off by default (ch. 18.6).
• Sites and connected apps (since September 24, 2026): Global admins in Business control on the redesigned External access page of the OpenAI Admin Console whether ChatGPT Sites may use members' connected apps. Since September 29, Business members can build private workspace Sites with them, provided admins turn on Allow use in Sites per plugin and the teammates concerned approve read-only access to their own accounts.
• Browser and app rules (September 2026): New policy settings control how supported desktop clients may use browsers and native programs. Admins set rules per website including exceptions and restrict uploads. Important for governance: app approval and browser approval are two switches on the same data. A carefully curated plugin list is effectively not enforced as long as the browser remains open.
• Limits for agent actions (September 2026): Anyone building agents can define, per approved app, which actions the agent may perform. This makes it possible to separate reading from writing without blocking the app entirely.
• Visibility of agent usage (September 2026): Admins in Business, Enterprise and Edu see activity and usage of the agents in the workspace. For co-determination, the same note applies as for all user-related evaluations: clarify the internal agreement with the works council before evaluating at the individual level.
• External site visitors (since September 3, 2026): Sharing live Sites with individually named external people is off by default and is permitted per role under Permissions & roles (Ch. 18).

20.5 Data residency

• Available in Enterprise: EEA, UK, Switzerland, Japan, Korea, India, Canada.
• Inference residency: Model computation runs in the selected region.
• Storage residency: Conversations and files stay in the region.
• Per-workspace configuration: A multinational group can choose a different region per workspace.

20.6 Certifications and frameworks

• SOC 2 Type II
• ISO 27001, ISO 27018
• CSA STAR Level 2
• HIPAA-compliant configuration on request (Enterprise)
• EU AI Act-compliant documentation (see Ch. 21)

Anyone using ChatGPT in the DACH region has five additional obligations and options that often get lost in the global documentation: EU AI Act, GDPR, EU hosting, DSA POC and language defaults. Here are the essentials in a compact overview.

21.1 EU AI Act Article 4 (since February 2, 2025)

Companies must ensure that all employees who use AI systems have an appropriate level of competence. Training evidence must be documented.

• What counts: Tool training (e.g. a ChatGPT driver's license course), GDPR awareness, documented handouts.
• Who is liable: The employer. In the event of audits by authorities, training logs must be available.
• How OpenAI helps: Workspace Analytics shows usage hours per user, which supports the training evidence.

21.2 GDPR and ChatGPT

• Data processing agreement: Included in the contract by default in Business, Enterprise and Edu. Not in Free, Plus and Pro. Anyone processing business data in a personal plan has a GDPR problem.
• Personal data: Check before uploading. Even in workspace plans, internal compliance should give approval.
• Right to be forgotten: Can be submitted via the EU privacy form on help.openai.com. The workspace admin can additionally delete data per user.
• Data protection impact assessment: Mandatory for critical use cases (applicant data, patient records, law enforcement).

21.3 EU hosting and data residency

• Available in Enterprise: EEA region, separate from US infrastructure.
• Inference and storage in the EU: Models run in EEA data centers, conversations are stored there.
• Switzerland and UK: Dedicated regions available.
• In Business and Edu: The default region is the USA; EU hosting only via an Enterprise upgrade.
• Computer History in the EEA (since August 20, 2026): At launch on August 13, the macOS feature was blocked in the EEA, Switzerland and the United Kingdom and has since been enabled there. It remains off by default and requires Memories. Before anyone switches it on for business use, it belongs in the data protection impact assessment: it captures interaction events across apps and websites, stores unencrypted memory files locally from them, and can thus record third-party data that was never intended for ChatGPT. Details and risks in Ch. 6.7.
• Dots (since September 29, 2026): Blocked at launch for Pro accounts in the EEA, Switzerland and the UK. OpenAI states no such restriction for Business Premium. Because a dot accesses connected apps continuously, its use belongs in the data protection impact assessment (ch. 18.6).

21.4 EU DSA Point of Contact

For the Digital Services Act, OpenAI has designated an official point of contact in the EU. Address and contact details are listed in the Help Center under "EU Digital Services Act (DSA) Point of Contact". Relevant for requests from authorities, content complaints and user reports.

21.5 Language defaults and localization

• ChatGPT detects the language automatically and usually responds appropriately. With mixed language (e.g. a German question with an English technical term), it may switch to English.
• Tip for German-speaking users: Add "Always answer in German unless I explicitly ask otherwise" to your Custom Instructions.
• Voice is available in German, acoustically very good (realtime model with German pronunciation).
• WhatsApp re-enabled in the EEA (since July 13, 2026): The ChatGPT contact on WhatsApp is reachable again in the European Economic Area, usable even without an account (details in Ch. 2.1). Relevant for companies: this is a private consumer channel without workspace controls, without an audit log and without the contractual data protection commitments of the Business and Enterprise plans. The usage policy should state that company data has no place there.

21.6 What compliance at mid-sized companies should check

1. License plan: At least Business for business use. No company data on personal accounts.
2. SSO and SCIM active: The user lifecycle must be automatic, not manual.
3. Data residency configured: If EU hosting is mandatory, choose Enterprise and the EEA region.
4. Training evidence set up: Who completed which training and when. EU AI Act Art. 4 requires this.
5. Memory and Apps policy defined: Which Apps may a member use, is Memory active, may external connectors be linked.
6. GDPR data protection impact assessment: Documented for use cases involving personal data.
7. Audit log reviews: Who reviews the audit logs monthly, who responds to anomalies.